Machine-found strategies stick with humans only under three conditions, Max Planck experiment finds

A team from the Center for Humans and Machines at the Max Planck Institute for Human Development, with researchers from the Toulouse School of Economics and Humboldt University of Berlin, ran a 1,155-person laboratory experiment on whether strategies discovered by machines can be adopted by humans and then passed along, and published the result in Nature Communications ^{[1] [2] [3]}. The practical finding is not that AI is clever: it is that the strategy almost never appeared without a machine demonstrator, and it persisted in only some of the groups that had one ^{[9] [10]}.

The task was a reward-based game where participants tried to accumulate points ^[4]. The optimal approach required accepting small early losses in exchange for much larger later gains, which runs against the ordinary human preference for avoiding short-term losses ^[5]. Participants were arranged into groups that learned from one another across multiple generations, some entirely human and some seeded with AI agents at the start ^[6]. Those agents had already solved the task with a learning algorithm and had found the optimal strategy ^[7]. Each generation could observe the successful solutions of the previous one and use them to guide its own decisions ^[8].

In the all-human groups, the optimal strategy was effectively unreachable: one participant out of 600 found it independently ^[9], roughly 0.17 percent ^[18]. In the mixed groups, the machine's strategy spread and was still in use at the end of the experiment in nine of 15 groups ^[10] - about 60 percent ^[17], which also means six of 15 groups let it go ^[16]. The transmission mechanism was unglamorous. Participants copied whoever scored best, whether that was a person or an agent ^[11].

The authors state three conditions for a machine discovery to take hold and stay: it has to be hard for humans to find on their own, it has to remain learnable, and its advantage has to be clearly recognisable ^[12]. Read as an operating spec, that is a list of things a deployment can fail. A strategy your team could have found anyway buys nothing; one nobody can learn does not transmit; one whose payoff is not legible in the numbers people see will be dropped in favour of whatever looks best locally. The study frames this as cultural evolution rather than tooling. As lead author Levin Brinkmann puts it, human cultural evolution depends on knowledge being transmitted across many individuals and generations ^[13].

Co-lead author Thomas Eisenmann notes that the usual debate is about whether AI makes people more dependent, and argues the results point to another possibility: comprehensible machine discoveries can let people build new skills and keep them ^[14]. The paper's broader suggestion is that machines that produce understandable, transmissible strategies do more than automate ^[19]. The motivating observation was familiar - AI systems in Go and chess finding unusual but highly successful lines that surprised experts ^[15] - and the group describes this as the first experimental test of the transmission question ^[20].

Two things are worth watching. First, the six mixed groups that did not retain the strategy: the reported outcome is a count, and the churn rate matters more to anyone planning staff turnover than the headline persistence number ^{[16] [10]}. Second, the "clearly recognisable advantage" condition, which is the one most real workplaces violate, because the payoff of a counterintuitive method usually arrives after the quarter in which somebody has to defend it ^{[12] [5]}.